Nano Banana 14 reference images vs Sume's 10 input_references

On Sume, Nano Banana 2 and Nano Banana Pro accept up to 10 input_references per request, not 14. Google's page says Gemini 3 image models can mix up to 14 reference images, and Sume's catalog code sets a ceiling of 10 for its edit-capable Nano Banana entries. ChatGPT Image 2.5 on Sume takes up to 16.

Google's number is from its image generation page; Sume's from Image generation and editing and the image catalog code, read 2026-09-30.

What are the limits side by side?

How do I send references on Sume?

Put them in the input_references array as image_url objects. The docs require reference URLs to be public HTTPS; localhost, private-network and non-HTTPS URLs are rejected before submission. A model whose input_references descriptor is {"min": 0, "max": 0} is text-to-image only and rejects references.

Why not just send 14?

Sending more than the model's maximum does not fit its range descriptor, so plan for the ceiling rather than the Google number. Read the input_references range from the discovery route for the exact model and use that as your cap.

What if I need more than 10 references?

Choose which references matter most, or use ChatGPT Image 2.5, which lists 16; see GPT Image 2.5 editing. For consistent characters on Nano Banana, see character consistency.